search

hawkular / hawkular-datamining

Real-time time series prediction library with standalone server
36 stars 22 forks source link

Computing prediction intervals of Exponential Smoothing models #83

Open amoralico opened 7 years ago

amoralico commented 7 years ago

I'm not sure whether the way predcition/confidecne intervals is correct. I think you must use the formulas that appear here and here (models ANN, AAN, AAA)

pavolloffay commented 7 years ago

There are more variants of these algorithms if I remember correctly. As a reference I used some literature from Rob J Hyndman and also forecast package from R.

amoralico commented 7 years ago

Yes of course. The general aproximation you use for prediction intervals : predictValue +- Z95% * sqrt(var(residuals)) is correct, but the variance (var) is different for each model (simple, double and triple). The last link I suggested (from Rob J Hyndman) shows how to compute these variances analitically, for the models you have implemented in the library (ANN = Simple, AAN = Double, AAA = Triple). I guess they are the Holt-Winters additive models.